Senior Software Engineer

Senior Software Engineer (Backend / DevOps / AI) - Sevan AI (Mumbai)
A foundational engineering role scaling a production AI accounting platform from one server to thousands of clients

About us
Sevan AI automates the most tedious part of bookkeeping: getting supplier bills out of inboxes and into the accounting system, accurately, without a human retyping them. Accountants forward or upload bills; our AI pipeline reads them (multi-model OCR with a verification layer), codes them to the right ledger accounts and tax treatment, and syncs them into QuickBooks Online or Xero for review. We are built by Orbit Accountants, an accounting firm, so the product is shaped daily by real bookkeepers processing real client books, not guesses about what accountants want.

The product works and is in production with paying clients. We are now scaling it from a founder-built system to a platform serving thousands of accounting clients, and you'd be one of the first engineers, co-owning that journey.

The Role
This role sits at the intersection of three disciplines, and it needs genuine depth in each: not familiarity, depth.

Backend: you'll own the document pipeline end to end: intake workers (email, upload, mobile), verification, and publishing to accounting systems. You'll also own the third-party integrations that surround it: QuickBooks Online, Xero, Gmail, Google Drive, with all the OAuth flows, token refresh, rate limits, and webhooks that come with vendor APIs.

DevOps: you'll turn a working single-server product into a real platform and own how it's built, shipped, and operated: environments, CI/CD, managed and backed-up data stores, containerization, monitoring and alerting. Most of it doesn't exist yet; you get to do it right. Scaling is not a side quest here; it is the job. You'll decide what the platform runs on, which technologies we adopt at each stage of growth, and which ones we deliberately defer.

AI engineering: you'll own the layer that actually reads and codes the bills: multi-model OCR orchestration, prompt and output-schema design, verification and confidence thresholds, evals against real documents, fallback strategies across providers, and per-document cost and latency budgets. This is the product; when the AI layer gets smarter, cheaper, or more reliable, every client feels it.
You'll also make the judgment calls that tie all three together: accuracy thresholds, retry semantics, and data-model decisions that touch customers' actual books.

Scaling the platform

- Evolve the architecture from a single server to a horizontally scalable platform: containerized services, a load-balanced app tier, and a worker fleet for document processing

- Introduce the right technology at the right time as volume grows: proper job queues (e.g. Celery or RQ with Redis), caching, object storage for documents, read replicas and query optimization on MySQL, cloud hosting (AWS/GCP or equivalent)

- Keep the AI layer fast, accurate, and affordable at volume: concurrency and batching for OCR calls, provider rate-limit management, fallback strategies across models, eval-driven accuracy tracking, and per-document cost tracking

- Build the observability to run at scale with a small team: metrics, tracing, structured logging, and alerting (e.g. Prometheus/Grafana, Sentry, or equivalents you'd advocate for)

- Design the security and data posture that holds up as client count grows: authentication and authorization, OAuth token and secrets handling, multi-tenant data isolation, backups and retention. This is customers' financial data, and it has to be right.

What we are looking for

- 5+ years building production backend systems in Python: FastAPI or similar, strong relational data modeling, and MySQL (or equivalent) tuned under real load. Our backend is Python end to end.

- You've personally run production systems. You've set up CI/CD from scratch, containerized and deployed services, migrated a live database, built the monitoring that caught the next incident, and been the one responsible when things broke at 2am. Not "worked at a company that had a platform team," but did it yourself.

- You've scaled a system beyond one machine. You know when to reach for queues, caching, read replicas, or horizontal scaling, have run workloads on AWS/GCP or equivalent, and can articulate which technology a system needs now versus at 10 the traffic.

- You've shipped LLM-powered systems to production. You're fluent in prompt and schema design, structured outputs, multi-model orchestration and fallbacks, and building eval and verification loops that catch model mistakes before customers do. You know how to manage cost, latency, and rate limits at volume, and you treat model output as untrusted input.

- Real third-party API integration experience, ideally with OAuth-heavy, imperfect vendor APIs. If you've integrated with accounting, payments, or banking APIs, tell us; Intuit or Xero experience is a jackpot.

- Fluent with AI-assisted development. We build with tools like Claude Code as a core part of the workflow. We are a small, senior team using AI for leverage, and we expect you to be excellent at specifying, reviewing, and owning AI-accelerated work.

- Security-minded by default: you think about authentication, token and secrets storage, and tenant isolation as part of the design, not as an afterthought.

- Comfortable with ambiguity and breadth. Some weeks are platform weeks, some are pipeline-debugging weeks, some involve digging into why a specific invoice was misread. You like owning outcomes, not lanes.

Nice to have: accounting or fintech domain knowledge; React/TypeScript or React Native familiarity (we have a web app and an Expo mobile app, though they're not your primary surface).

This job is probably not for you if...
- You want a well-defined lane with a platform team handling everything below your service
- You're deep in one of backend, DevOps, or AI engineering but only conversational in the other two; every part of this role leans on all three
- Your reflex for a two-server problem is Kubernetes and a service mesh; we scale deliberately, adopting what the workload proves we need
- You treat the AI layer as a black box someone else tunes; here, the model calls are product code you own
- You prefer greenfield work over making a real, revenue-generating system excellent

Our stack
Python (FastAPI) • MySQL • React + TypeScript • Expo/React Native • Gemini and OpenAI APIs • QuickBooks Online + Xero APIs • Gmail API • Claude Code in daily use.

As we scale, you'll help choose what we add next: expect Docker, Redis-backed job queues, managed cloud infrastructure, and a real observability stack to be on the near-term roadmap, with you holding the pen on those decisions.